Model Context Protocol concept

What is MCP?

Reference

What is MCP? MCP, or Model Context Protocol, is an open protocol that gives AI applications a standard way to exchange context and request capabilities from external services. An application connects through MCP clients to MCP servers, which can offer data, reusable prompts, or actions. MCP sets rules for that exchange; it does not determine how an AI application uses the information it receives.

Reviewed 2026-10-09

A shared way for AI apps to connect

MCP is a protocol: a shared set of rules for exchanging messages between an AI application and programs that provide data or capabilities. The application is called the host. It creates an MCP client to communicate with each server; a server is a program that provides context or capabilities. The client maintains the connection between its host and that server.

A server might offer data for the application to read, a reusable prompt, or a tool the application can ask it to run. This gives application makers a common interface for integrations. MCP describes the protocol for context exchange; it does not prescribe how an AI application uses a language model or manages the context it receives. The parent topic covers the protocol’s parts in more detail.

How the pieces communicate

In the MCP specification revision dated 2026-07-28, the host uses a client to communicate with a server. The messages use JSON-RPC, a format for requests and responses. The client can discover which protocol versions and capabilities a server supports before trying to use them.

Servers can offer three main kinds of capability: tools, which perform actions; resources, which provide data; and prompts, which are reusable message templates. The protocol separates these messages from the transport, the channel that carries them. The 2026-07-28 specification defines standard transports including standard input and output for communication between local processes, and Streamable HTTP for messages sent over HTTP. The same protocol message rules apply across transports.

How MCP relates to APIs and AI models

MCP sits between an AI application and programs that provide data or actions. It is not itself an AI model, database, or external service. An MCP tool can perform an API call, but the API call is an action the tool carries out; MCP defines how the host, client, and server exchange protocol messages.

The protocol also does not decide what an application does with information it receives. The host is the AI application that coordinates clients, while the client and server handle the MCP connection. This distinction helps locate MCP among its neighbours: it standardizes one part of an integration, while the application and connected services supply the rest. See MCP servers versus APIs for that comparison.

What MCP does not guarantee

MCP does not mean every server offers every capability. Servers declare the capabilities they support, and clients can discover those capabilities. A connection may offer tools, resources, prompts, or a subset of them.

MCP also does not make a tool’s effects harmless or decide whether a user should approve an action. The specification calls out data access and code execution as security considerations, and says the protocol itself cannot enforce its security principles. The host application and its implementation have responsibilities for consent and access controls. Treat a server’s available actions as something the application and user need to understand.

Questions people ask

What does MCP stand for?

MCP stands for Model Context Protocol.

Does MCP connect an AI model directly to a service?

MCP defines communication between a host application, its client, and a server. The host is the AI application; MCP does not dictate how that application uses its language model.

Is MCP the same thing as an API?

No. An MCP tool can make an API call, but MCP is the protocol for exchanging messages between the host, client, and server; the API call is an action a tool may perform.

Does every MCP server provide tools?

No. Servers can offer tools, resources, prompts, or a combination, and clients can discover a server’s supported capabilities.

Does MCP itself guarantee that a tool call is safe?

No. The specification identifies security considerations around data access and code execution, and says the protocol itself cannot enforce its security principles.

Learn more

What is the Model Context Protocol (MCP)?
The official introduction gives a concise explanation of MCP and examples of the systems it connects.
Architecture overview
The official architecture guide explains the host, client, server, protocol layers, transports, and core capabilities.

What changed

Sources

Part of Model Context Protocol: Connection roles and server capabilities Assess what a connection can do